Treat Noise as Domain Shift: Noise Feature Disentanglement for Underwater Perception and Maritime Surveys in Side-Scan Sonar Images

In underwater perception and maritime surveys, due to the scarcity of training data and perturbation of speckle noise, the detection performance of underwater objects in side-scan sonar (SSS) images is limited. To address these problems, we proposed a noise feature disentanglement you only look once (YOLO) (NFD-YOLO) by combining noise-agnostic features learning and attention mechanism. First, we rethink the speckle noise by treating it as the domain shift between the training dataset and real-measured SSS images and build a domain generalization-based (DG-based) underwater object detection framework. Then, we extend YOLOv5 with a feature manipulation module, a noise-agnostic subnetwork, and an auxiliary noise-biased subnetwork for noise features disentanglement, more biases toward noise-agnostic features and less reliance on noise-biased features in underwater object detection, respectively. Finally, the ACmix attention module is introduced for a more powerful learning capacity and attention to the object areas based on a small dataset. According to the experiment results, the proposed NFD-YOLO achieved 75.1% mean average precision (mAP) in the test domain, which increased by 7.5% than YOLOv5, and 75.7% +/- 0.4% mAP and 77.5% +/- 1.6% mAP for different speckle noise distributions and transfer directions, respectively, which verified its generalization ability and robustness for speckle noise; therefore, the proposed method can mitigate the effects of speckle noise and provides a new thought to address the speckle noise in underwater object detection with a small dataset, which is of significance and benefits for underwater perception and maritime surveys.

Automated summary

This paper proposes a method to improve the detection performance of underwater objects by disentangling noise features. The method treats speckle noise as the domain shift between training data and real-measured images, and incorporates feature manipulation, noise-agnostic subnetwork, and auxiliary noise-biased subnetwork to handle noise features effectively. Additionally, the ACmix attention module is introduced to enhance learning capacity and focus on object areas.